Improved metabolomic data-based prediction of depressive symptoms using nonlinear machine learning with feature...

Area of Science:

• Psychiatric epidemiology
• Computational biology
• Metabolomics

Background:

• Depressive symptoms pose a significant public health challenge.
• Existing metabolite-based prediction algorithms have limitations.
• Accurate prediction of psychiatric phenotypes is crucial for early intervention.

Purpose of the Study:

• To develop a novel prediction model for depressive symptoms using nonlinear feature selection machine learning.
• To apply the Hilbert-Schmidt independence criterion least absolute shrinkage and selection operator (HSIC Lasso) algorithm to a large-scale metabolomic dataset.
• To improve the accuracy of predicting depressive symptoms based on metabolomic profiles.

Main Methods:

• Utilized a population-based dataset of 897 subjects from communities affected by the Great East Japan Earthquake.
• Employed 306 metabolite features derived from nuclear magnetic resonance and mass spectrometry.
• Developed and evaluated prediction models using nested fivefold cross-validation.

Main Results:

• The HSIC Lasso model demonstrated superior predictive power compared to Lasso, support vector machine, partial least squares, random forest, and neural network models.
• Key contributing metabolites included L-leucine, 3-hydroxyisobutyrate, and gamma-linolenyl carnitine.
• The model successfully integrated nonlinear feature selection for enhanced prediction accuracy.

Conclusions:

• The HSIC Lasso-based prediction model significantly improves the prediction of depressive symptoms from metabolome data.
• Identified specific risk metabolites associated with depressive symptoms using nonlinear statistics in the Japanese population.
• Further research should explore the generalizability of these findings across different ethnicities.